What is MQL5 CodeGraph, and why does AI need evidence-backed graphs?

An AI agent can read a few MQL5 files quickly. The hard part starts when a repository contains many includes, classes, event handlers, and cross-file calls. Important context gets separated, while a plausible inference can quietly be repeated as fact.

MQL5 CodeGraph provides orientation through a static graph snapshot. It reads .mq5 and .mqh files, records declarations, includes, calls, and standard MetaTrader runtime relationships, and retains relationship type, direction, origin, and source location.

Four questions a useful graph should answer

• Where is this symbol declared, and who calls it?
• Which path can connect OnTick to risk-management logic?
• What may be affected if a shared function changes?
• Which result came from source, and which result is inference?

The last question matters most. A graph edge does not prove that terminal execution followed that path. Static analysis also does not replace MetaEditor compilation, backtesting, or runtime observation.

One kernel, several surfaces

The canonical snapshot is interpreted by the Intelligence Kernel. The CLI, local dashboard, and MCP adapter project the same query, context, impact, and directed-path semantics. That gives an agent and its operator a result they can cross-check instead of an answer with no trail.

Limits worth stating plainly

A tolerant tokenizer is not a complete MQL5 compiler. Complex macros, dynamic dispatch, runtime-generated names, or incomplete source may produce partial results. Diagnostics are part of the evidence, not noise to hide.

MQL5 CodeGraph is designed for trusted local repositories and a single-user loop. It does not prove that an EA is safe, strategically correct, or profitable.

Within that boundary, a code graph is a practical map. It does not drive for the developer, but it helps both people and AI understand where they are.